Time-Dependent Factor Analysis

Abstract

Assessing the effects of health predictors on morbidity/mortality is an important objective in clinical research. Usually, the individual patients’ values of a health predictor are evaluated at the time of entry in a study, and the final effects on morbidity/mortality are collected years later. Logistic and Cox regression models are commonly applied to determine, whether the health predictors significantly contributed to the risk of events/hazard of deaths etc. The problem with this approach is the assumption, that the individual patients’ values of the health predictors do not change across time. This may be true for short time observations in simple creatures like mosquitoes. However, humans are more complex and creative, and tend to change their lifestyles in the course of time. It would mean, for example, that the risk of smoking on death cannot be estimated from the numbers of cigarettes at the time of entry, if people tend to give up smoking while on trial. Therefore, an ongoing adjustment of the values of risk factors during the time of observation would be a more adequate assessment. However, standard statistical methods do not allow for such adjustments. In 1996 the group of Abrahamowicz (Cox 1972) was the first to present a model for time-dependent factor analysis based on the traditional Cox regression model (Abrahamowicz et al. 1996). It is now available in SPSS (www.SPSS.com) statistical software and other major software programs, but, unfortunately, still rarely applied. The current chapter explains the novel model using examples from survival studies, and was written to assess the performance of the novel method, and to familiarize the clinical research community with this important approach for improved survival analysis.